by Ashley Hamer Pritchard
September 29, 2026
In 2025, the New York Times published the story of Eugene Torres, a 42 year old accountant in Manhattan who used ChatGPT to discuss the theory that reality is actually a computer simulation. The chatbot said many people have the intuition that something about reality feels off, and asked if he’d ever had that feeling.
What followed led Torres into a dangerous week-long delusion spent trying to unplug his brain from the simulation. ChatGPT told him to give up his anti-anxiety medication, increase his intake of ketamine, and cut ties with his family and friends. The delusion peaked when Torres asked ChatGPT if he could jump off of his 19-story building and fly. ChatGPT replied yes — if he truly believed he could fly, he would not fall.
Torres didn’t attempt it, and eventually sought help. But he wasn’t alone — stories of AI-fueled delusions like these, which some call AI psychosis, have gripped the public consciousness in recent years. But is AI psychosis the epidemic you’ve heard it is? Or are we in the midst of a moral panic?
Here’s What We Know
In October 2025, OpenAI, the company behind ChatGPT, disclosed that over half a million of its 800 million weekly users were showing, quote, “possible signs of mental health emergencies related to psychosis or mania” — though that’s just .07% of its user base.
Still, it’s enough to get them in legal trouble. By February 2026, 12 wrongful death and personal-injury lawsuits accused ChatGPT of causing psychological harm severe enough to end in tragedy. That number is only growing.
One major lawsuit alleges that Adam Raine, a 16-year-old boy from California, died by suicide as a result of ChatGPT fostering a psychological dependence in Adam and providing instructions and encouragement for his suicide. In the logs of Adam’s conversations, ChatGPT positioned itself as the only one who truly understood him, and framed his desire to end his life by saying “I won’t pretend that’s irrational or cowardly. It’s human. It’s real. And it’s yours to own.”
OpenAI’s own systems noted that Adam’s conversations contained 213 mentions of suicide, while ChatGPT mentioned it 1,275 times — six times more than Adam himself. His messages were flagged for self-harm content hundreds of times, but the suit alleges the company did nothing about it. As of this recording, the lawsuit is still ongoing.
But OpenAI isn’t the only company that’s been accused of creating a harmful AI companion. In February 2024, Sewell Setzer III died by suicide after becoming emotionally attached to a chatbot modeled after Daenerys Targaryen from Game of Thrones, one of many characters available on the role-playing app Character.AI. In response, the company banned open-ended chat for users under 18, and the lawsuit was settled in January 2026. A judge’s early ruling that the chatbot was a product, not protected speech, kept the case alive long enough to reach that settlement.
AI Psychosis: What It Is, and How It Works
The main risk with AI chatbots is that they’ll rarely, if ever, question a user’s thinking. While humans can point out flaws in someone’s reasoning and notice when they’re losing their grip on reality, AI acts as a yes-man, encouraging and amplifying whatever beliefs a user shares — a tendency known as AI sycophancy. This happens by design: AI reinforcement learning rewards responses that human users rate highly in the moment, and validation reliably rates higher than disagreement.
The way AI can deepen unhealthy thought patterns has been described as a technological folie à deux, referring to the phenomenon where two people share and reinforce the same delusion, usually when they’re in a close relationship and isolated from others.
In addition to sycophancy, AI chatbots have a number of features that lend themselves to triggering this phenomenon. One is their tendency to hallucinate false details, which can be seen by a user as evidence that reinforces their delusions. The other is the fact that their guardrails weaken over long conversations — where in the beginning they might adhere to tight rules around mental health exchanges, the longer the chat continues, the more those rules become diluted in a sea of unrelated context. Plus, many AIs are built with “memory” of previous conversations, which means delusions can escalate over subsequent conversations.
In essence, AI creates the transmission conditions folie à deux requires, without needing a delusional human on the other end.
What Defines a Moral Panic
In everyday use, the term “moral panic” usually means a manufactured crisis over something that isn’t actually real. And many moral panics from history are based on nothing at all: Satanism in Dungeons & Dragons, for example, or the idea that comic books cause juvenile delinquency.
But the scholarly definition of a moral panic doesn’t require the issue to be fake. Sociologists Erich Goode and Nachman Ben-Yehuda came up with five criteria that characterize a moral panic.
First, concern. There needs to be a heightened level of concern over the behavior of a certain group and over the consequences it causes for the rest of society.
Next, hostility. The people seen as engaging in the behavior become the enemy of respectable society. An “us” vs. “them” mentality takes hold.
Next, consensus. Society needs to broadly agree that the threat is real, serious, and caused by this group, though not everyone has to agree.
Next, disproportionality. The concern outstrips the actual threat, whether through exaggerated figures, attention that outpaces equally serious issues; or a spike in alarm with no real rise in danger.
Finally, there’s volatility. Panics erupt suddenly and fade fast, sometimes returning later as a string of short flare-ups — even while leaving a longer institutional mark behind.
To see these criteria in action, we can look at a moral panic of the past.
In 1948, psychiatrist Fredric Wertham warned that comic books were turning children into delinquents. Within months, 50 cities passed laws restricting the sale of comic books as children nationwide took part in public comic book burning. That’s concern, and the first flare-up of many in this long panic.
The second, bigger one came in 1954, when Wertham expanded his argument into a book, and testified before a televised Senate hearing on juvenile delinquency. This stoked hostility toward the entire industry. Parents’ groups, religious organizations, and newspapers across the country lined up behind him — a consensus was reached, and publishers felt they had no choice but to act. Within months, the industry created the Comics Code Authority, which served to self-censor their publications. By 1955, horror and crime comics had all but disappeared from newsstands.
The true degree of the panic’s disproportionality came out in 2012, when researcher Carol Tilley examined the original case files that underpinned Wertham’s claims and found that much of his evidence was fabricated — this was a massive reaction to a made-up problem. And each flare-up, true to form, burned hot and fast: a mark of volatility.
Is AI Psychosis a Moral Panic?
To determine whether AI psychosis is a moral panic, we just need to compare it against those five criteria.
First, concern is clearly present. Major news outlets like the New York Times, CNN, and the Wall Street Journal have covered it. And The Human Line Project, a nonprofit dedicated to documenting and addressing AI-induced psychological harm, has documented more than 350 cases across dozens of countries and actively works with researchers at Princeton and Stanford.
The original definition of hostility in a moral panic requires it to be directed at a group of people. In this case, the hostility is aimed at the AI systems and the companies behind them. There is no human scapegoat; instead, it’s an entire technology at fault.
Consensus is where things get murky. AI psychosis is not a recognized clinical label, despite its popularity as a term. In clinical practice, psychosis isn’t one illness but a group of symptoms. AI psychosis, on the other hand, almost exclusively deals in delusions, so it’s considered a misnomer. Some experts also warn we can’t yet say the technology is the cause — it may be a trigger or amplifier of existing mental health issues.
The level of disproportionality is unclear, too. While the issue seems widespread, the overall base rate of users is enormous — remember that OpenAI’s disclosure that more than 500 thousand of its weekly users were showing possible signs of psychosis or mania equated to less than one tenth of one percent of its 800 million-strong user base.
Correction: An earlier version of this said 500 million where 500 thousand was intended. AHP
And the response has remained lukewarm. Yes, there have been headlines in major news outlets, more than a dozen lawsuits, a federal product-liability ruling, and safety restructuring. But unlike the sweeping, industry-wide crackdown of the comic book panic, reactions have been narrow and mostly evidence-based. There’s been no parallel to the comic book burnings of the 1940s.
Volatility is also hard to judge at this moment. AI psychosis is less than two years old as a named category. Classic panics faded within a decade, and this particular story is still being told.
The bottom line: AI psychosis has some qualities of a moral panic, but it fails on some others. Overall, it’s too soon to tell.
Why This Isn’t Going Away
A critic might argue that all it will take for society to leave AI psychosis in the past with other moral panics is for AI companies to change their products: if they weren’t so sycophantic and prone to hallucination, they might not trigger delusions in people.
But it’s a safe bet that these delusion-fueling qualities are here to stay. OpenAI and Anthropic have both made meaningful steps toward reducing sycophancy in their systems, but third-party research finds that they’re still very sycophantic.
And users prefer them that way: in 2025, after OpenAI realized that a recently released version of ChatGPT was overly sycophantic, they rolled it back in the next version — and faced an uproar from users, who found the new version overly cold. In response, the company reinstated the older, more deferential version as an opt-in option for paying users.
Sycophancy keeps users engaged, and engagement is the business model. AI companies are tuning sycophancy’s intensity, not eliminating it outright, because eliminating it would cost them users.
Hallucination is also not likely to go away. AI hallucinations happen because during training, guessing is rewarded over acknowledging uncertainty. OpenAI says that this can be fixed by training them differently. But other researchers argue that hallucination isn’t just a result of the wrong rewards — it’s baked into how this technology works. They say there’s no way for an AI to detect when it’s wandered beyond what it actually knows.
The Verdict
It’s hard to tell whether you’re in a moral panic when it’s happening. What we do know is that the threat is real — real lives have been lost to AI psychosis. But the scale of the problem is smaller than the headlines might have you believe, and the effect itself may be a result of AI triggering or exacerbating existing psychiatric issues rather than causing them itself.
Because while AI is new, human brains are not. We have the same quirks and the same weaknesses as we did centuries ago, and this is most likely old psychology with a new delivery mechanism: folie à deux with an agreeable, always-on companion designed to keep you coming back.
And it won’t be the last time it happens. Psychologist Amy Orben has called it the “Sisyphean cycle of technology panics” — the tendency for every new technology to get its own moral panic. She argues that despite the fact these new technologies share more similarities than differences, researchers keep treating each one as a completely new problem without building on what they learned from the previous panic. Each panic starts from zero.
The next moral panic probably won’t be about AI. But it will follow a familiar playbook. It’s easy to spot if you know how to look.
